This is so true! Working on Vercel CDN is challenging and often intense, but never, ever boring. I've used more computer science in this role than any previous job.
Ten million deployments a day makes the invisible plumbing a lot more interesting.
Love seeing the engineering behind something that usually looks like one deploy button.
This was fun to work on. It was our third CDN project where we designed and shipped JSONL based formats that are both technically efficient and operationally simple at scale.
The Vercel CDN makes over 80 million routing decisions per second.
We cut P99 metadata lookup latency by 91%, helping requests reach your pages faster and speeding up deployments along the way ↓
vercel.com/blog/how-we-cut-c…
Starting August 10, Vercel's CDN will stop stripping the 𝚂𝚎𝚛𝚟𝚎𝚛-𝚃𝚒𝚖𝚒𝚗𝚐 response header and pass it through to the browser.
Report backend metrics like database query time, and they show up in the network panel and the Performance API.
vercel.com/changelog/server-…
“All are invited to participate in a unified fast to express gratitude for religious liberty and to pray that it be strengthened throughout the world,” wrote Church President @OaksDallinH and his counselors in the First Presidency.
thechurchnews.com/members/20…
Remember everyone, LLMs are amazing math models that can make extremely correct *looking* code.
By pure coincidence, these are sometimes right and the more guardrails and feedback you give it, the more likely it's actually correct.
But finding error is *very* hard, by design.
Imagine if you had one of those genies that would grant wishes, but always in the most subtly wrong way while technically doing exactly what you asked.
Then imagine you asked it to write code that looked correct and big free.
The genie smirks and says "as you wish"
The resulting code is amazing and upon initial inspection looks perfect. Even running it appears to just work on the first try.
But you said it needs to *look* correct, not that it had to *be* correct.
Be careful out there and wish wisely.
When Opus 4.5 came out, it was a one-way door to a new way of engineering. Agents now do most of our coding.
Knowing the inherent flaws and over-confidence of LLMs, we sent a clear message to our teams. Vibing and mission-critical infrastructure don’t go together.
We’re sharing some of our early internal guidance in how we’re “agenting responsibly”, prioritizing security, durability, and availability at all times.
vercel.com/blog/agent-respon…
Bulk redirects on Vercel now scale to 1 million per project with near-constant lookup time.
We shipped the simplest thing first, profiled it in production, and iterated from there ↓
vercel.com/blog/scaling-redi…
Hey @theo, what's stopping you from adding code sandbox to t3.chat? Is the context required just too long or sandbox cost concerns or something else? I'd love a client-only wasm interpreter that can run some LLM based toy language as a tool call.
I've been iterating on a new encoding format that's LLM friendly. It's a very interesting challenge because tokens are not bytes and especially challenging when you consider that the LLM needs to be able to comprehend the custom format. This makes competing with JSON hard.
What's really exciting is that the pretty-printed version is actually less bytes than the compact form of JSON. It's more tokens than compact JSON because tokenizers optimize for JSON and the pretty-printed whitespace hurts, but still impressive nonetheless.
"O remember, remember, ...; yea, remember that there is no other way nor means whereby man can be saved, only through the atoning blood of Jesus Christ, who shall come; yea, remember that he cometh to redeem the world."